2023 年 31 巻 p. 34-44
In group discussions, each participant has his/her own opinion, and if necessary, tries to persuade other participants to reach an agreement. Therefore, persuasiveness is an important skill for communicating with others. Based on this motivation, this study aims to estimate the persuasiveness of group discussion participants. First, human annotators rated the level of persuasiveness of each participant in group discussions among four people. We then created multimodal and multiparty models for estimating persuasiveness of a participant using speech, language, and visual (head pose) features using gated recurrent unit (GRU)-based neural network. In our experiment, in estimating the highest persuasive participant among the group, the performance of the proposed method was 76% in accuracy. In the binary classification task for estimating high or low persuasiveness participants among the group, the performance of the best performing multimodal multiparty model achieved 80% in accuracy. The experimental results show that multimodal models are better than unimodal models, and multiparty features contribute to improving the model performance. As an application of the proposed method, we implemented a meeting browser with persuasiveness visualization. The level of persuasiveness is visually indicated using the background color of each participant's timeline. Finally, we conducted a user study for our meeting browser, and found that the persuasiveness visualization helps the subjects grasp the flow of the discussion in a shorter time compared to a browser without persuasiveness visualization.
